miHoYo Unveils Gamer AI for Game Control

The research team's vision-language-action framework plays Honkai: Star Rail using screen input and keyboard-and-mouse actions.

Updated on Nov. 22, 2026 in Artificial Intelligence

Isometric editorial illustration of a plain keyboard and mouse linked by a teal light band to a stack of stepped translucent frames on cream, suggesting an AI system acting on screen observations.
miHoYo's research team unveiled Gamer, an AI framework that plays Honkai: Star Rail using only screen input and keyboard-and-mouse actions, with a preprint submitted November 13 and revised November 18. AI Illustration. Upload story photo >

Chinese game developer miHoYo's research team described Gamer, a vision-language-action framework that maps screen observations and instructions to keyboard and mouse actions. The accompanying preprint was submitted November 13, 2026 and revised November 18, 2026.

Why it matters

The work shows an AI agent operating a commercial game from screenshots alone, without internal game interfaces, a step toward general-purpose agents that control software and virtual environments. The authors frame the framework as applicable across conventional graphical interfaces and three-dimensional game worlds.

The main experimental model builds on the 8-billion-parameter Qwen3-VL-8B base and was trained through continued pretraining on gameplay trajectories, instruction tuning and online reinforcement learning. Independent verification of the author-reported performance figures is not yet available, and the Honkai: Star Rail training data remains withheld.

The players

miHoYo

A Chinese video game developer best known for the role-playing titles Genshin Impact and Honkai: Star Rail.

The details

The model takes screenshots, task information, observation history and memory summaries as input, using sparse reasoning to preserve task progress as earlier interactions leave its active context. A cloud-edge serving system called Streaming-vLLM is described alongside the framework, and checkpoints from the continued-pretraining scaling study have been released.

Timeline

  1. November 13, 2026: Preprint first submitted.

  2. November 18, 2026: Preprint revised.

  3. November 21, 2026: Chinese account detailing the work published.

The Tech Race

The work extends the Qwen3-VL-8B vision-language base into a game-control agent, adding to a wave of AI systems that learn to operate interfaces from visual input rather than structured APIs. It positions Chinese developers alongside Western labs pursuing general-purpose computer-use agents.

Gamers may eventually see AI companions or automated assistants that play alongside them, though the research is not yet a released product. The released checkpoints let other researchers test similar game-control agents without access to the withheld Honkai: Star Rail data.

The takeaway

Agents that act through screens and keyboard-and-mouse input rather than internal APIs are becoming a practical research direction. Results from developer-run evaluations are worth weighing until independent testing confirms them.

Further reading

Explore more on Artificial Intelligence developments in game-playing and agent research.

More information

Read the Source article on Gamer research for the author-reported evaluation details.

Source note: This article includes information reported by Pandaily.